Understanding the Effects of Optimal Combination of Spectral Bands on Deep Learning Model Predictions: A Case Study
Md Abul Ehsan Bhuiyan1, Chandi Witharana1, Anna K Liljedahl2,3
1Department of Natural Resources and the Environment, University of Connecticut, Storrs, CT 06269, USA.
Journal of Imaging
|August 30, 2021
Summary
Selecting optimal multispectral bands is crucial for deep learning models in remote sensing. This study shows that specific band combinations significantly impact the accuracy of detecting Arctic ice-wedge polygons using Mask RCNN.
Area of Science:
- Remote Sensing
- Geospatial Analysis
- Artificial Intelligence in Earth Science
Background:
- Deep learning (DL) convolutional neural networks (CNNs) are increasingly used for very high spatial resolution (VHSR) satellite image analysis.
- Multispectral (MS) imagery, including infrared bands, is vital for earth science applications like geo object detection and classification.
- Understanding how MS band statistics influence DLCNN predictions is essential for optimizing model performance.
Purpose of the Study:
- To investigate the extent to which multispectral (MS) band statistics influence deep learning convolutional neural network (DLCNN) model predictions.
- To evaluate the impact of different three-band combinations on DLCNN model performance for detecting Arctic ice-wedge polygons (IWPs).
Main Methods:
- Utilized the Mask RCNN DLCNN architecture for detecting ice-wedge polygons (IWPs) in eight-band Worldview-02 VHSR satellite imagery.
- Designed a systematic experiment involving five cohorts of three-band combinations, assessing spectral variability and its impact on model predictions.
- Employed statistical measures to quantify model performance, including F1 score, random error, and systematic error.
Main Results:
- Achieved high F1 scores (0.89-0.95) for IWP detection using specific three-band combinations, notably coastal blue, blue, green (1,2,3) and green, yellow, red (3,4,5).
- The coastal blue, blue, green (1,2,3) band combination demonstrated low random (0.17-0.19) and systematic (0.20-0.21) errors in IWP mapping.
- Model prediction accuracy was found to be significantly influenced by the selection of input multispectral bands.
Conclusions:
- The careful selection of optimal spectral bands is critical for enhancing DLCNN prediction accuracy, especially when restricted to three channels.
- Image statistics of input multispectral bands play a significant role in the performance of DLCNN models for geospatial applications.
- This research underscores the importance of band selection strategies in remote sensing for accurate feature detection and classification.
